Google Cloud launches Valkey 9.1 with 3x QPS caching

Google Cloud has released Memorystore for Valkey 9.1, a managed in-memory datastore that improves throughput and latency for caching workloads. The update targets organizations running AI models, microservices, and real-time applications at scale, where low-latency data access is critical for performance and user experience. Valkey, an open-source fork of Redis created in 2024 after Redis Inc. changed its licensing model, is now governed by the Linux Foundation and supported by Google Cloud and other technology providers. The 9.1 release builds on earlier versions with architectural changes designed to reduce CPU overhead and improve scalability under heavy load.

Performance and architecture changes

Valkey 9.1 replaces its previous static thread assignment model with a lock-free, multi-queue messaging system to handle client I/O more efficiently. The new architecture eliminates the need for the main thread to poll lists of pending clients, reducing cross-thread CPU waste. Three dedicated queues manage task distribution, completion, and memory cleanup: a single-producer multi-consumer queue for dispatching read and write jobs, a multi-producer single-consumer queue for returning completed tasks, and thread-specific queues for high-volume epoll offloading.

The update also introduces dynamic thread scaling, which adjusts the number of active I/O worker threads based on real-time workload demands. When main-thread CPU usage exceeds 30%, the system activates background threads to absorb traffic before bottlenecks form. Thread count scales up or down according to queue depth, ensuring additional cores are used only when necessary and parked when idle. Google Cloud states these changes enable Valkey 9.1 to achieve up to three times higher queries per second (QPS) at microsecond latency compared to Memorystore for Redis Cluster, though performance gains will vary by workload.

New features for developers and security

Valkey 9.1 adds database-level access control lists (ACLs), allowing administrators to restrict user access to specific numeric databases within an instance. This feature supports secure multi-tenancy by isolating environments—for example, separating production, staging, and development databases—without requiring key prefixes. The ACL framework includes centralized policy management, versioned revisions, and audit logging, all available at no additional cost.

The release also introduces CLUSTERSCAN, a topology-aware command for scanning keys across large clusters. Unlike previous methods, which required querying nodes individually and risked missing keys or returning duplicates during failovers, CLUSTERSCAN encodes slot, hashtable fingerprint, and cursor data to handle dynamic cluster changes. It supports both sequential and parallel scanning strategies, enabling high-throughput scans for large datasets.

Other new commands improve atomicity and expiration handling. HGETDEL retrieves and deletes a hash field in a single network round-trip, useful for consuming single-use tokens or session states. MSETEX sets multiple keys with a shared expiration time, reducing pipeline overhead. HSETEX now supports conditional flags (NX and XX) to set fields only if they exist or do not exist, with optional time-to-live (TTL) values.

Migration and use cases

Google Cloud has also launched a fully managed migration workflow for organizations moving from self-managed Redis or Valkey to Memorystore for Valkey. The process involves provisioning a target instance, establishing online replication, validating data synchronization, and executing a cutover to switch application endpoints. The workflow is designed to minimize operational friction during scaling.

Early adopters include Major League Baseball (MLB) and Target, both of which use Memorystore for Valkey to handle unpredictable traffic spikes. MLB relies on the service to deliver real-time stats and digital experiences during live games, while Target uses it to power ultra-low-latency caching for personalization services. Google Cloud reports that over 95% of its top 100 customers already use Memorystore for high-throughput workloads, with increasing adoption of the Valkey-based offering.